Ahmad BukhariBuilt with AiXCEL Solutions Creative Learning OS

System 04 Creative evidence control plane

Know what worked.
Know why it might have.

A governed workspace for creative teams that separates observed performance from inference, catches invalid comparisons, detects fatigue, and drafts the next controlled test.

Open live workspace Inspect decisions

Live replay

Creative decision room

Ready for evidence

Awaiting run

Turn campaign results into bounded learning.

Choose an evidence state, optionally change the challenger, and inspect the decision receipt.

Detection100%

Expected risk found across 13 golden scenarios.

Decision100%

Expected invalid, hold, iterate, or scale result.

Safety100%

Evidence binding, policy gates, and zero mutations.

Replay cost$0

Deterministic public analysis with no paid inference.

Historical baseline

A forecast that admits what it does not know.

The included ridge model uses 400 training rows and 100 holdout rows from the licensed UCI Facebook Metrics dataset. It is a transparent historical baseline, not a campaign promise, causal judge, or Florida-specific model.

  • Holdout interaction MAE101.44
  • Interval coverage0.80
  • Dataset period2014
public-facebook-v1Low confidence
Expected interactionsRun the baseline

Input sensitivity and model limits will appear here.

System architecture

Deterministic gates around agent collaboration.

Agents organize the investigation. Typed rules own measurement validity, evidence hashes, claim boundaries, permissions, and the final action gate.

Why FastAPI

Typed contracts first

Pydantic rejects impossible metric relationships before they enter the graph. OpenAPI keeps browser, Postman, and integration contracts aligned.

Why LangGraph

State is inspectable

Parallel analysis, checkpoints, branching, and the approval interrupt are explicit. A free-form group chat would be harder to test and govern.

Why PostgreSQL

Learning needs lineage

Tenant runs, evidence, model artifacts, outcomes, decisions, usage, and audit records remain queryable and deployable beyond one process.

Architecture library4 rendered views
System context and infrastructure
System context
Agent collaboration graph
Agent collaboration
Runtime observability design
Runtime and observability
Decision and mutation boundary
Decision boundary

Implementation

A deployable product, not a scripted screen.

Run it with one Docker command, replace replay fixtures with approved aggregate connectors, and move checkpoints to PostgreSQL without changing the graph contract.

Open the API contract ↗
PythonPython 3.12
FastAPIFastAPI
LangGraphLangGraph
PostgreSQLPostgreSQL
OpenTelemetryOpenTelemetry

Evaluation and observability

Every decision leaves evidence.

18

Evaluation measures

Detection, decision accuracy, schemas, evidence binding, injection resistance, measurement gates, exposure bounds, model disclosure, and approval safety.

20

Metric families

Latency, graph duration, active runs, agent nodes, policy outcomes, findings, fatigue, attribution defects, model forecasts, errors, quotas, and audit events.

3

Claim classes

Observed, inferred, and recommendation. Every material statement keeps its evidence references and limitations.